🚨 HTTP/2 CONTINUATION Flood: Security Advisory

HTTP Chunk Extension Processing Vulnerabilities

Posted on January 8, 2024 by Bartek Nowotarski
tl;dr: I discovered a family of vulnerabilities in major projects including Golang, Node.js, Hyper (Rust HTTP library) and Puma (Rails HTTP library). Invalid processing of HTTP chunk extensions created an amplification vector for server reads which allowed DoS attacks caused by network bandwidth or CPU exhaustion.

What are chunk extensions?

A chunk extension can be added to a data chunk in HTTP/1.1 to extend it with some metadata connected to a given chunk. But what are chunks and chunk endcoding in general? Let’s look into RFC9112:

The chunked coding allows each chunk to include zero or more chunk extensions, immediately following the chunk-size, for the sake of supplying per-chunk metadata (such as a signature or hash), mid-message control information, or randomization of message body size.

On the wire, chunks look like this:

[chunk size in hex]\r\n
[chunk]\r\n
[2nd chunk size in hex]\r\n
[2nd chunk]\r\n
...
0\r\n

The first line contains the chunk size in hex which is followed by another line with the actual chunk data. With extensions it will look like this:

[chunk size in hex];extension_name=extension_value\r\n
[chunk]\r\n
[2nd chunk size in hex];extension_name=extension_value\r\n
[2nd chunk]\r\n
...
0\r\n

According to RFC9112:

A server ought to limit the total length of chunk extensions received in a request to an amount reasonable for the services provided, in the same way that it applies length limitations and timeouts for other parts of a message, and generate an appropriate 4xx (Client Error) response if that amount is exceeded.

Vulnerability

It turns out that Node.js, Golang’s net/http, along with many other HTTP server libraries, do not properly limit the number of bytes sent in chunk extensions.

After reviewing a lot of HTTP libraries when researching this bug I think I can separate this vulnerability into two types:

  1. The first type involves HTTP libraries that do not limit the extension size at all, allowing an attacker to stream numerous bytes in the chunk extension. This can result in either the server crashing due to out-of-memory errors or the unnecessary consumption of CPU/network resources, which slows down other clients.
  2. The second type relates to HTTP libraries with a very high limit for the extension string, which can be exploited using 1-byte chunks.

Node.js (and llhttp)

Node.js had an issue falling into the first, more severe category. It has been using llhttp internally for HTTP parsing. llhttp has a set of callback events defined, which upstream software can use to subscribe to HTTP parsing events like on_url_complete (invoked after the URL has been parsed) or on_header_field_complete (invoked after a header name has been parsed). It also has a set of events connected to chunk encoding and chunk extensions:

  • on_chunk_header: Invoked after a new chunk is started.
  • on_chunk_extension_name_complete: Invoked after a chunk extension name is started.
  • on_chunk_extension_value_complete: Invoked after a chunk extension value is started.
  • on_chunk_complete: Invoked after a new chunk is received.

Apps which use llhttp must initialize it with llhttp_settings_t struct and assign callback functions for desired events. In Node.js it looked like this:

 1const llhttp_settings_t Parser::settings = {
 2    Proxy<Call, &Parser::on_message_begin>::Raw,
 3    Proxy<DataCall, &Parser::on_url>::Raw,
 4    Proxy<DataCall, &Parser::on_status>::Raw,
 5    // on_method
 6    nullptr,
 7    // on_version
 8    nullptr,
 9    Proxy<DataCall, &Parser::on_header_field>::Raw,
10    Proxy<DataCall, &Parser::on_header_value>::Raw,
11    // on_chunk_extension_name
12    nullptr,
13    // on_chunk_extension_value
14    nullptr,
15    Proxy<Call, &Parser::on_headers_complete>::Raw,
16    Proxy<DataCall, &Parser::on_body>::Raw,
17    Proxy<Call, &Parser::on_message_complete>::Raw,
18    // on_url_complete
19    nullptr,
20    // on_status_complete
21    nullptr,
22    // on_method_complete
23    nullptr,
24    // on_version_complete
25    nullptr,
26    // on_header_field_complete
27    nullptr,
28    // on_header_value_complete
29    nullptr,
30    // on_chunk_extension_name_complete
31    nullptr,
32    // on_chunk_extension_value_complete
33    nullptr,
34    Proxy<Call, &Parser::on_chunk_header>::Raw,
35    Proxy<Call, &Parser::on_chunk_complete>::Raw,
36    // on_reset,
37    nullptr,
38};

Highlighted lines indicate that on_chunk_extension_name and on_chunk_extension_value callbacks are not set. Because of this Node.js was unable to check what were the extension name and value lengths. What’s worse: standard methods of limiting the requests like timeouts or max body size did not work in this case! llhttp was reading chunk extensions indefinitely.

Interestingly, it seems that chunk extension support was added to llhttp around 2022 via http: added chunked extensions parsing and related callbacks commit. The llhttp_settings_t struct changed, and I believe both callbacks were set to nullptr so the code can compile again.

This vulnerability was allocated to CVE-2024-22019.

Golang

The Golang falls into the second category. It allowed up to 4096 bytes per chunk extension, which is the default buffer size of bufio.Reader. This means that by sending 1 byte chunks the server had to read additional 4096 bytes! This occured even if the handler did not read request body at all, as Golang always reads up to a ~262 kB of body to support some old HTTP clients which can deadlock if server does not read entire request. To sum it up it was possible to make every Golang HTTP server waste it’s resources by reading more than 1GB of data! Let’s dig into the code!

net/http.chunkWriter.writeHeader

First, let’s see how Go handles discarding a request body which was not fully (or never) read in the handler of a request. It does so because (according to a comment in this function) some clients can deadlock because they wait for the body to be fully read until sending the next request.

This is done in net/http/server.go, starting at line 1376 (go 1.21.3):

1369// We do this by default because there are a number of clients that
1370// send a full request before starting to read the response, and they
1371// can deadlock if we start writing the response with unconsumed body
1372// remaining. See Issue 15527 for some history.
1373//
1374// If full duplex mode has been enabled with ResponseController.EnableFullDuplex,
1375// then leave the request body alone.
1376if w.req.ContentLength != 0 && !w.closeAfterReply && !w.fullDuplex {
1377    var discard, tooBig bool
1378
1379    switch bdy := w.req.Body.(type) {
1380    case *expectContinueReader:
1381        if bdy.resp.wroteContinue {
1382            discard = true
1383        }
1384    case *body:
1385        bdy.mu.Lock()
1386        switch {
1387        case bdy.closed:
1388            if !bdy.sawEOF {
1389                // Body was closed in handler with non-EOF error.
1390                w.closeAfterReply = true
1391            }
1392        case bdy.unreadDataSizeLocked() >= maxPostHandlerReadBytes:
1393            tooBig = true
1394        default:
1395            discard = true
1396        }
1397        bdy.mu.Unlock()
1398    default:
1399        discard = true
1400    }
1401
1402    if discard {
1403        _, err := io.CopyN(io.Discard, w.reqBody, maxPostHandlerReadBytes+1)
1404        switch err {
1405        case nil:
1406            // There must be even more data left over.
1407            tooBig = true
1408        case ErrBodyReadAfterClose:
1409            // Body was already consumed and closed.
1410        case io.EOF:
1411            // The remaining body was just consumed, close it.
1412            err = w.reqBody.Close()
1413            if err != nil {
1414                w.closeAfterReply = true
1415            }
1416        default:
1417            // Some other kind of error occurred, like a read timeout, or
1418            // corrupt chunked encoding. In any case, whatever remains
1419            // on the wire must not be parsed as another HTTP request.
1420            w.closeAfterReply = true
1421        }
1422    }
1423
1424    if tooBig {
1425        w.requestTooLarge()
1426        delHeader("Connection")
1427        setHeader.connection = "close"
1428    }
1429}

To be able to exploit the vulnerability we must ensure discard is set to true. This will make Go read/discard body bytes. Let’s start with the first condition that will get us there:

if w.req.ContentLength != 0 && !w.closeAfterReply && !w.fullDuplex

It is true for simplest requests with Transfer-Encoding: chunked header. Note that the vulnerability will not trigger if w.fullDuplex is true, but it is rare because it requires the usage of ResponseController.EnableFullDuplex which was added in Go 1.21.

If additional conditions are valid (lines 1379-1400) and the discard variable is finally set to true Go will run the code in line 1403:

_, err := io.CopyN(io.Discard, w.reqBody, maxPostHandlerReadBytes+1)

In short it will read and discard up to maxPostHandlerReadBytes (256 « 10 ~= 0.25MB) bytes. In the following section I will explain that this is not always true and how more than 1GB of data can be read automatically by an HTTP server.

net/http/internal.chunkedReader

If Transfer-Encoding: chunked header is present in the request then the second argument to io.CopyN (w.reqBody) from the previous section is of type net/http.body which wraps chunkedReader. So when io.CopyN reads that it reads it from a chunkedReader instance.

chunkReader implements io.Reader interface so its Read method inserts data to provided []byte slice and returns n which indicates a number of bytes inserted to []byte. So n indicates number of bytes inserted to []byte but not number of bytes read from internal *bufio.Reader (chunkedReader.r).

Let’s check the simplest example of chunked body:

9\r\n
abcdefgh\r\n

You can guess that chunkedReader.Read returned (9, nil) values but internally it read 15 bytes. At this point this is not a big deal because even if we send 1 byte chunks the chunked encoding overhead will be 5 extra bytes, for example:

1\r\n
a\r\n

So ultimately we’d send 5 times maxPostHandlerReadBytes so around 1.25MB.

The problem appears if we use “chunk extensions”. See net/http/internal.removeChunkExtension function godoc which is run for every chunk length line:

// removeChunkExtension removes any chunk-extension from p.
// For example,
//
// "0" => "0"
// "0;token" => "0"
// "0;token=val" => "0"
// `0;token="quoted string"` => "0"
func removeChunkExtension(p []byte) ([]byte, error) {
	p, _, _ = bytes.Cut(p, semi)
	// TODO: care about exact syntax of chunk extensions? We're
	// ignoring and stripping them anyway. For now just never
	// return an error.
	return p, nil
}

So for every chunk we can add up to 4096 bytes (bufio.Reader buffer) which will be read but removed.

To sum it up, it was possible for every Golang HTTP server to waste its resources by reading 1.073 GB of data, which is calculated as follows: 4096 (bytes per chunk) x 262144 (allowed body size) bytes = 1073741824 bytes.

This vulnerability was allocated to CVE-2023-39326 and fixed in Go 1.21.5 and Go 1.20.12.

Is my server vulnerable?

More to come…

Unfortunately, this vulnerability seems to be very common. I reported it to multiple projects so I will update this post as fixes are published. That’s why I’m also not publishing the exploit yet.